Deep Learning of Neuromuscular and Visuomotor Control of a Biomimetic Simulated Humanoid

Masaki Nakada, Tao Zhou, Honglin Chen, Arjun Lakshmipathy, Demetri Terzopoulos

Abstract: We present a biomimetic framework for human neuromuscular and visuomotor control that promises to be of value to researchers developing humanoid robots. Our framework features a biomechanically simulated human musculoskeletal model, actuated by numerous skeletal muscles, with realistic eyes driven by extraocular and intraocular muscles, whose optic organs refract light, and whose retinas have many nonuniformly distributed photoreceptors. The humanoid's visuomotor control system comprises 24 trained deep neural networks (DNNs)-10 DNNs in its vision subsystem and 14 DNNs in its motor subsystem-plus an additional 4 trained shallow neural networks (SNNs) that control the irises and lenses of the eyes. Of the motor DNNs, a pair control the extraocular muscles, 6 per eye, responsible for eye movements, 2 control the 216 neck muscles of the cervicocephalic biomechanical complex, producing natural head movements, 2 control the 443 core muscles of the torso, and 2 control each limb; i.e., the 29 muscles of each arm and 39 muscles of each leg. Directly from the foveated retinal photoreceptor responses, a pair of foveation DNNs drive eye, head, and torso movements, while 8 limb vision DNNs extract the visual information needed to direct arm and leg actions. By synthesizing its own training data, our humanoid automatically learns efficient, online, active visuomotor control of its eyes, head, torso, and limbs in order to perform nontrivial tasks involving the foveation and visual pursuit of moving target objects coupled with visually-guided limb-reaching actions to intercept them. We also demonstrate that it can balance itself in an upright stance, take steps, and perform certain simulated sports activities.